Deep Learning for Logo Recognition

In this project we present a method for logo recognition based on deep learning. Our recognition pipeline is composed of a logo region proposal followed by a Convolutional Neural Network (CNN) specifically trained for logo classification.
Experiments are carried out on both the FlickrLogos-32 database and our extended Logos-32plus dataset.

Our classification model and extended dataset are made available for research purposes (link). If you use any of this material, please cite:

@article{bianco2017deep,
  title={Deep learning for logo recognition},
  author={Simone Bianco and Marco Buzzelli and Davide Mazzini and Raimondo Schettini},
  journal={Neurocomputing},
  volume={245},
  pages={23--30},
  year={2017},
  issn={0925-2312},
  doi={http://dx.doi.org/10.1016/j.neucom.2017.03.051},
  url={http://www.sciencedirect.com/science/article/pii/S0925231217305660}
 }
 
@inproceedings{bianco2015logo,
  title={Logo recognition using cnn features},
  author={Bianco, Simone and Buzzelli, Marco and Mazzini, Davide and Schettini, Raimondo},
  booktitle={International Conference on Image Analysis and Processing},
  pages={438--448},
  year={2015},
  organization={Springer}
 }

Datasets

FlickrLogos-32 (link) dataset is a publicly-available collection of photos showing 32 different logo brands. It is meant for the evaluation of logo retrieval and multi-class logo detection/recognition systems on real-world images.

Example images for each of the 32 classes of the FlickrLogos-32 dataset

Logos-32plus (link) dataset is an expansion of the train-set of FlickrLogos-32. It has the same classes of objects as its counterpart but a larger cardinality (12312 instances).
We collected this new dataset with the aim of taking into account a larger set of real imaging conditions and transformations that may occur in uncontrolled acquisitions.
The dataset contains on average 400 examples per class, with each image including one or multiple instances of the same class.

Graphical comparison of the distribution of the 32 logo classes between FlickLogos-32 and our augmented Logos-32plus dataset

To ensure a high variability of the new dataset and to avoid any overlap with the existing one, we performed a semi-automatic check for duplicate images within the Logos-32plus dataset itself and with the FlickrLogos-32 dataset.

Results

The following table reports methods trained on either FlickrLogos-32 or on the combination of both FlickrLogos-32 and Logos-32plus.
The test set is always the official test set of FlickrLogos-32.
Results refer to recognition performance, and are computed according to the standard evaluation protocol defined in fl_eval_classification.py.

Method Train data Precision Recall F1 Accuracy
BoW SIFT [1] FL32 0.991 0.784 0.875 0.941
BoW SIFT + SP + SynQE [1] FL32 0.994 0.826 0.902 N/A
Romberg et al [2] FL32 0.981 0.610 0.752 N/A
Revaud et al. [3] FL32 >0.980 0.726 0.834÷0.841 N/A
Romberg et al. [1] FL32 0.999 0.832 0.908 N/A
Bianco et al. 2015 [4] FL32 0.909 0.845 0.876 0.884
Bianco et al. 2015 + Q.Exp [4] FL32 0.971 0.629 0.763 0.904
Eggert et al. [5] FL32 0.996 0.786 0.879 0.846
Oliveira et al. [6] FL32 0.955 0.908 0.931 N/A
DeepLogo [7] FL32 N/A N/A N/A 0.896
Bianco et al. 2017 [8]
(This work)		 FL32 0.976 0.676 0.799 0.910
FL32, L32+ 0.989 0.906 0.946 0.958

Researchers are highly invited to submit their own results to for inclusion in this table.
Submissions should follow this example format, containing one class and confidence for each entry in the FlickrLogos-32 test set.

References

[1] S. Romberg, R. Lienhart, Bundle min-hashing for logo recognition, in: Proceedings of the 3rd ACM conference on International conference on multimedia retrieval, ACM, 2013, pp. 113–120.
[2] S. Romberg, L. G. Pueyo, R. Lienhart, R. Van Zwol, Scalable logo recognition in real-world images, in: Proceedings of the 1st ACM International Conference on Multimedia Retrieval, ACM, 2011, p. 25.
[3] J. Revaud, M. Douze, C. Schmid, Correlation-based burstiness for logo retrieval, in: Proceedings of the 20th ACM international conference on Multimedia, ACM, 2012, pp. 965–968.
[4] S. Bianco, M. Buzzelli, D. Mazzini, R. Schettini, Logo recognition using cnn features, in: Image Analysis and ProcessingICIAP 2015, Springer, 2015, pp. 438–448.
[5] C. Eggert, A. Winschel, R. Lienhart, On the benefit of synthetic data for company logo detection, in: Proceedings of the 23rd Annual ACM Conference on Multimedia Conference, ACM, 2015, pp. 1283–1286.
[6] G. Oliveira, X. Frazão, A. Pimentel, B. Ribeiro, Automatic graphic logo detection via fast region-based convolutional networks, in: Neural Networks (IJCNN), 2016 International Joint Conference on, IEEE, 2016, pp. 985–991.
[7] F. N. Iandola, A. Shen, P. Gao, K. Keutzer, Deeplogo: Hitting logo recognition with the deep neural network hammer, arXiv preprint arXiv:1510.02131.
[8] S. Bianco, M. Buzzelli, D. Mazzini, R. Schettini, Deep Learning for Logo Recognition, Neurocomputing 245, 23–30 (2017), http://dx.doi.org/10.1016/j.neucom.2017.03.051

Publications